How do pin an ML job to a specific runtime image version and python version?

[calleo]

Hi!

I am using the decorator below to start an ML job.

@jobs.remote(
    compute_pool=COMPUTE_POOL,
    stage_name=JOB_STAGE,
    database=ML_DATABASE_ENV,
    schema=JOBS_SCHEMA,
    runtime_environment='???',
)

I am trying to pin the runtime version to specific container runtime, and python version. But both doesn't seem to be possible?

I have tried 2.6.0, which resolves to Python 3.10 (my project uses 3.12).
I have also tried with {"pythonVersion": "3.12"}, which gives me the latest runtime version, with python 3.12. Problem is that latest might change, including the installed package version which I'd like to avoid.

Right now I have reverted to supplying the complete image name, but that will only work as long as we are using CPU compute pools: "/snowflake/images/snowflake_images/container_runtime/cpu_x86_64:2.6.0-py312".

How can I pin the environment to a container version and python version at the same time?

[sfc-gh-kunli]

AFAIK there is no supported way to specify both runtime version and Python version simultaneously through runtime_environment. The documented API accepts either:

• A version tag (e.g., "2.6.0") — which resolves server-side but doesn't let you control the Python version
• A full image URL (e.g., "/snowflake/images/.../cpu_x86_64:2.6.0-py312") — which pins everything

The workaround you can use is passing the full image path directly, which pins both:

# CPU pools
runtime_environment="/snowflake/images/snowflake_images/container_runtime/cpu_x86_64:2.6.0-py312"

# GPU pools
runtime_environment="/snowflake/images/snowflake_images/container_runtime/gpu_x86_64:2.6.0-py312"